Stochasma 0.5.0-rc.1

Stochasma 0.5.0-rc.1 is the release candidate for the generic conditioning layer.

Highlights:

• representation-agnostic conditioning protocol;
• conditioned predictor composition;
• classifier-free guidance primitives;
• conditioning-aware training batches;
• shared private validation and randomness layers;
• categorical classifier-free-guidance boundary validation;
• regression coverage across DDPM, DDIM, categorical sampling, training, RNG semantics, and paclet loading.

Stochasma 0.4.0

Stochasma 0.4.0 extends the general-purpose Wolfram Language diffusion toolkit.

Highlights:

• canonical D3PM predicted-x0 categorical reverse sampling;
• validation-once sampler hot paths;
• encoder-backed latent diffusion training;
• selectable DDPM/DDIM latent sampling;
• separated production and test code;
• clean paclet-loading and public-API integration tests.

Stochasma 0.3.0

Scalar categorical diffusion primitives and reverse sampling for the Wolfram Language.

Includes:

• uniform categorical transition kernels;
• arbitrary row-stochastic categorical transition sampling;
• validated categorical transition schedules and cumulative kernels;
• time-indexed categorical forward diffusion;
• exact categorical reverse posteriors;
• canonical joint-marginalized predicted-x0 reverse probabilities;
• categorical single-step reverse sampling;
• full scalar categorical reverse sampling;
• automatic, locally seeded, and explicit-noise randomness;
• reverse trajectories from caller-supplied x_T to x_0.

The 0.3 categorical reverse sampler is scalar-only. Array/tensor categorical reverse sampling and terminal-prior helpers are not part of this release.

Stochasma 0.2.0

Model-facing utilities and extended Gaussian diffusion sampling for the
Wolfram Language.

Includes:

• deterministic sinusoidal time embeddings;
• NetChain and NetGraph predictor adapters;
• reproducible diffusion training batches;
• deterministic and stochastic DDIM sampling;
• full and subsampled DDIM timestep trajectories;
• explicit, seeded, and automatic randomness semantics;
• DDIM/DDPM consistency tests.

Stochasma 0.1.0

Initial stable Gaussian DDPM core for the Wolfram Language.

Includes:

• linear and cosine beta schedules;
• canonical DDPM coefficients;
• closed-form forward diffusion;
• clean-sample reconstruction;
• posterior mean and variance;
• stochastic and explicit-noise reverse steps;
• epsilon-prediction training primitives;
• predictor-driven DDPM sampling;
• explicit RNG semantics for automatic, seeded, and explicit-noise modes;
• numerical validation of diffusion schedules;
• headless tests and a synthetic Gaussian example.